Brain–computer interfaces

Why it matters

Amyotrophic lateral sclerosis, brainstem stroke and spinal cord injury can leave a person unable to move or speak while the brain still works. A brain–computer interface offers a channel for messages and commands that does not depend on muscles ^[1].

What it is

A brain–computer interface (BCI) is a non-muscular channel for sending messages and commands to the outside world, built from recordings of brain activity ^[1].

• Signals used include slow cortical potentials, P300 potentials and mu or beta rhythms recorded from the scalp, and the activity of cortical neurons recorded by implanted electrodes ^[1].
• The signals are translated in real time into commands that operate a computer display or another device ^[1].

How it works

The user has to encode commands in the signals and the system has to derive the commands from them, so the user and the BCI adapt to each other, both at first and continually ^[1].

Brain–computer interface signal chainBrain signals are recorded, translated in real time into commands, and the commands operate a device; the user and the system adapt to each other.

Step through

• Excitatory synapse (filled arrowhead)
• Inhibitory synapse (bar)
• Modulatory (open circle)
• Signal or data flow, not a synapse (dashed)
• Midline crossing (decussation)

Text description of this diagram

1. Brain signals are recorded, translated in real time into commands, and the commands operate a device; the user and the system adapt to each other.
2. User’s intent (midline, Brain) to Signals (midline, Recording); signal or data flow (not a synapse).
3. Signals (midline, Recording) to Translation (midline, Translation); signal or data flow (not a synapse).
4. Translation (midline, Translation) to Device (midline, Output); signal or data flow (not a synapse); commands.
5. Device (midline, Output) to User’s intent (midline, Brain); signal or data flow (not a synapse); feedback: both adapt.

The numbers

Speed is compared in bits. The information transfer rate counts how many bits each selection carries, given the number of targets and the accuracy ^{[3, 1]}:

Information transfer rate (Wolpaw)^{[3, 1]}

$$B = \log_2 N + P \log_2 P + (1 - P) \log_2 \frac{1 - P}{N - 1}$$

How many bits each selection of a brain–computer interface carries, given the number of possible targets and the chance of choosing the right one. Multiply by selections per minute for bits per minute.

Variables of Information transfer rate (Wolpaw)

Symbol	Meaning	Unit
$N$	number of targets	dimensionless
$P$	accuracy	%
$B$	bits per selection	bit/selection

Valid when

• Assumes every target is equally likely and errors spread evenly over the wrong targets
• Accuracy below chance (1/N) gives misleading values
• Says nothing about how long each selection takes; that comes in through selections per minute

Worked example^[3]

Four targets chosen correctly 90% of the time carry about 1.37 bits per selection, against 2 bits if it were always right.

N = 4 , P = 90 % → B = 1.37 bit/selection

number of targets (dimensionless)accuracy (%)

bits per selection: 1.373 bit/selection

In 2002 the best interfaces reached maximum rates of 10 to 25 bit/min^[1]. An implanted interface later decoded attempted handwriting at 90 characters per minute^[9] with 94.1 % raw accuracy online^[9].

In theatre and clinic

Implanted interfaces have been tested in small trials. Two people with long-standing tetraplegia used signals decoded from motor cortex neurons to control a robotic arm and hand for reach and grasp ^[4].

A brain–spine interface links cortical recordings to stimulation of the spinal cord regions that produce walking; one person with chronic tetraplegia used it to stand and walk in community settings ^[5].

Frontier

Each entry shows its evidence tier and what it does not show ^{[6, 7, 8]}.

• T1T1 peer-reviewed human studyTyping by imagined handwriting2021-05 · as of 2026-10-02 · trial

  An intracortical interface decoded attempted handwriting from motor cortex with a recurrent neural network. The participant, whose hand was paralysed by spinal cord injury, typed 90 characters per minute with 94.1% raw accuracy online.

  What it does not show: One participant with implanted arrays; it restores communication, not hand movement.

  Sources and details
• T1T1 peer-reviewed human studySpeech to text at 62 words per minute2023-08 · as of 2026-10-02 · trial

  Intracortical arrays recorded spiking activity while a participant with ALS attempted to speak. Word error rate was 9.1% on a 50-word vocabulary and 23.8% on a 125,000-word vocabulary, at 62 words per minute.

  What it does not show: One participant; with a large vocabulary about one word in four was wrong.

  Sources and details
• T1T1 peer-reviewed human studyText, voice and a speaking avatar from the speech cortex2023-08 · as of 2026-10-02 · trial

  High-density surface recordings of the speech cortex were decoded into text, synthesised speech and facial-avatar movement. Text decoding ran at a median 78 words per minute with a median word error rate of 25%, after less than two weeks of training.

  What it does not show: One participant; error rates remain far above natural speech.

  Sources and details
• T1T1 peer-reviewed human studyA speech neuroprosthesis that calibrates quickly2024-08 · as of 2026-10-02 · trial

  Four microelectrode arrays (256 electrodes) in the left ventral precentral gyrus of a man with ALS reached 99.6% accuracy with a 50-word vocabulary on the first day, 90.2% with a 125,000-word vocabulary on the second day, and sustained 97.5% accuracy over 8.4 months; he used it to converse at about 32 words per minute for more than 248 hours.

  What it does not show: One participant in a clinical trial; it needs brain surgery and is not an approved device.

  Sources and details

• T1T1 peer-reviewed human studyA brain-spine interface lets a man with tetraplegia walk again2023-05 · as of 2026-10-02 · trial

  Two 64-electrode implants over the sensorimotor cortex read the intention to move a hip, knee or ankle; a decoder turns it into stimulation of the lumbosacral cord through an implanted paddle lead. The participant, with an incomplete cervical injury from ten years earlier, stood, walked and climbed stairs; the system stayed reliable for a year including use at home, and he regained some walking with crutches even with it switched off.

  What it does not show: One participant, with an incomplete injury and earlier stimulation training. It does not repair the cord, and it is not an approved treatment.

  Sources and details

The AI connection

Decoders are fitted models: they learn a map from neural activity to intended movement or text. A recurrent neural network decoded attempted handwriting in one participant ^[9].

Common misconceptions

Check yourself

Four targets, chosen correctly 90% of the time: how many bits per selection?

About 1.37 bits, by the Wolpaw formula, against 2 bits if every selection were right ^[3].

What did the brain–spine interface connect?

Implanted cortical recordings to epidural stimulation of the spinal cord regions involved in walking ^[5].

Read next

• Primary motor cortex, the usual recording site.
• BCI decoder lab and digital bridge lab.